AI/ML Engineer · B.Tech CSE, GLA University

Krishna Gupta

I ship production AI — NLP matching, privacy-aware RAG, and applied ML — with live apps you can open, not just notebooks.

4 Live apps
8.38 CGPA / 10
3 Certifications
NLP · CV · RAG Focus
View selected work Resume ↗
kg434580@gmail.com

Selected work

Projects

AI/ML engineer from Moradabad, India. B.Tech CSE (AI & ML) at GLA University (CGPA 8.38/10). Selected work: live NLP screening, a local-first PII/RAG auditor, clinical ML, and YOLOv8 drain-blockage detection. Azure AI-900 & AZ-900; Google Cybersecurity Professional Certificate.

Loading preview…
AI Resume & Job Matching preview Live
Full-Stack AI Live

AI Resume & Job Matching

Hiring matcher that combines TF-IDF overlap with SpaCy similarity, then uses Gemini to pull implicit skills from job descriptions. Each run logs Precision, Recall, and F1 — not a single opaque score.

Objective
Replace keyword ATS filters with a measurable NLP + LLM pipeline.
Stack
TF-IDF + SpaCy + Gemini 1.5 Flash
Backend
Flask MVC + Firebase Firestore
  • Dual matcher: TF-IDF keywords plus SpaCy cosine similarity
  • Gemini skill extraction for requirements buried in JD prose
  • Blind hiring mode, RBAC, CSRF, and per-IP rate limits
PythonFlaskSpaCyGeminiFirestoreVercel
Visit Website ↗ Source Code ↗
Loading preview…
RiskLens preview Live
Privacy & Compliance Live

RiskLens

In-memory PII auditor: regex, Luhn/Verhoeff checksums, and spaCy NER, scored with a CDF risk curve. Offline TF-IDF RAG for Q&A; Gemini/OpenAI stay opt-in so documents stay local by default.

Objective
Local-first compliance scan with optional RAG — no default API leak.
Stack
Regex + spaCy NER + TF-IDF RAG
Backend
Flask in-memory sessions + Docker
  • Zero-disk pipeline — files stay in memory; nothing written to disk
  • Luhn + Verhoeff to cut false positives on cards and Aadhaar
  • Offline RAG with optional LLM summaries (metadata only unless opted in)
PythonFlaskspaCyRAGTesseractDocker
Visit Website ↗ Source Code ↗
Loading preview…
Diabetes Prediction System preview Live
Applied ML Live

Diabetes Prediction System

End-to-end clinical baseline on the PIMA Indians dataset: impute, scale, compare Logistic Regression / Random Forest / SVM, then serve a Flask form. Held-out accuracy 70.8% — a deployable pipeline, not a SOTA claim.

Objective
Ship a complete sklearn pipeline, not a notebook-only demo.
Dataset
PIMA Indians Diabetes Dataset
Result
Accuracy 70.8% · Precision 60% · Recall 50% · F1 54.6%
  • Train → evaluate → serve with Joblib on a simple clinical form
  • Zero-as-missing imputation and StandardScaler before model compare
  • Held-out 70.8% accuracy / 50% recall — recall favored to miss fewer cases
PythonFlaskScikit-learnPandasJoblib
Visit Website ↗ Source Code ↗
FloodSentinel

In Development — preview available on GitHub

In Development
Computer Vision In Development

FloodSentinel

YOLOv8n trained on storm-drain CCTV to flag sewage blockages and estimate occlusion. Unseen-test mAP@0.5 0.685. Map dashboard and sensor fusion are designed; the repo today is the detector and occlusion tooling.

Objective
Catch drain blockages before flash floods, not after.
Stack
Fine-tuned YOLOv8n (Ultralytics)
Result
Test mAP@0.5 0.685 on held-out CCTV frames
  • YOLOv8n on 703/198/98 train/val/test drain images (Roboflow)
  • Unseen test: mAP@0.5 0.685, precision 0.631, recall 0.680
  • Occlusion scoring scripts; municipal map UI not shipped yet
PythonYOLOv8PyTorchUltralyticsOpenCV
View on GitHub ↗
Loading preview…
AssetFlow preview Live
Hackathon ERP Live

AssetFlow

Enterprise asset tracker from the 24-hour Odoo Hackathon 2026 (team of 4). Seven lifecycle states, overlap-blocked bookings, and four RBAC roles on Firebase.

Objective
Replace spreadsheet asset tracking with role-based workflows.
Event
Odoo Hackathon 2026 (24 hours)
Team
Ansh Yadav (lead), Krishna Gupta, Abhay Kumar, Adarsh Katara
  • 24-hour sprint with 4 teammates; live on Firebase Hosting
  • 7 asset states and rules that block double-allocation
  • 4-role RBAC with Google sign-in and a maintenance approval flow
ReactTypeScriptFirebaseTailwind
Visit Website ↗ Source Code ↗
TrueProfile AI

In Development — preview available on GitHub

In Development
Multi-Modal AI In Development

TrueProfile AI

Five-signal fake-account scorer (ArcFace, 2D-FFT, XGBoost behavior, Random Forest metadata, NetworkX) fused with logistic regression. Numbers below are from the project's own 2,500-profile benchmark — not a public leaderboard.

Objective
Score fake accounts from several signals, not one heuristic.
Stack
ArcFace + FFT + XGBoost + RF + NetworkX
Result
In-repo 5-fold CV on 2,500 synthetic/benchmark profiles
  • Five signals fused with a logistic meta-model and a review queue
  • Internal 5-fold CV: fusion 99.4% acc on a 2,500-profile set (treat as in-repo, not SOTA)
  • FastAPI + Alembic/PostgreSQL, rate limits, and API-key auth
PythonFastAPIXGBoostArcFacePostgreSQL
View on GitHub ↗
J.A.R.V.I.S

In Development — preview available on GitHub

In Development
Voice AI Agent In Development

J.A.R.V.I.S

Voice-first agent with Gemini 2.0 Flash (search grounding) and Groq Llama 3.3 failover in Express. MongoDB session memory; wake-word and double-clap DSP. React HUD is still in planning.

Objective
Hands-free assistant with explicit model failover.
Stack
Gemini 2.0 Flash + Groq Llama 3.3 in Express
Memory
MongoDB profile/session nodes
  • Dual-LLM router: Gemini 2.0 Flash primary, Groq Llama 3.3 fallback
  • Wake-word + double-clap DSP for hands-free triggers
  • MongoDB memory nodes; cinematic HUD designed, not built
Node.jsExpressMongoDBGeminiGroq
View on GitHub ↗
DRISHTI

Early Stage — preview available on GitHub

Early Stage
Multimodal Deep Learning Early Stage

DRISHTI

Research pipeline for image+text sentiment on MVSA (~19,600 pairs) across 7 planned languages. Data loading, transforms, tokenization, and W&B/Hydra scaffolding are tested — encoders and training are not implemented yet.

Objective
Catch sarcasm that lives in caption + image, not text alone.
Dataset
MVSA (~19,600 pairs) — pipeline tested, model untrained
Tracking
Weights & Biases + Hydra
  • ~19,600 MVSA pairs with a tested PyTorch Dataset / DataLoader
  • XLM-RoBERTa + ResNet50 / ViT transforms sized; fusion not trained
  • Hydra configs and W&B ready before the first training run
PyTorchHuggingFaceXLM-RoBERTaHydraW&B
View on GitHub ↗

Capabilities

Skills

Tools and technologies I use to ship production AI systems.

Languages

PythonJavaJavaScriptTypeScriptCSQL

Machine Learning

Scikit-learnXGBoostPyTorchHuggingFaceYOLOv8TensorFlowPandasNumPy

NLP & LLMs

SpaCyRAGGeminiTransformersXLM-RoBERTaPrompting

Web & Full-Stack

FlaskFastAPIReactNode.jsExpressFirebasePostgreSQLMongoDB

Cloud & Tools

Microsoft AzureVercelDockerGitW&BPytest
LeetCode ↗
100+ problems solved
Live apps
4 deployments on Vercel and Firebase
Certifications
Azure AZ-900, AI-900 & Google Cybersecurity
Leadership
Former Technical Head, Saturangle Club

Education & Credentials

Education

Undergraduate specializing in AI & ML. CGPA 8.38/10. Building production NLP, RAG, and computer-vision systems alongside coursework.

  • CGPA 8.38/10 in B.Tech CSE — AI & ML
  • Live apps in NLP matching, PII/RAG audit, and clinical ML
  • CV: YOLOv8 drain-blockage detection (FloodSentinel, mAP@0.5 0.685)

Contributed to the club’s debating and public-speaking activities, organized debate events and discussions, and helped coordinate club initiatives and student participation.

  • Organized and coordinated debate events and competitions
  • Facilitated discussions and public-speaking activities
  • Helped manage club activities and student participation

Microsoft Azure Fundamentals (AZ-900) and Azure AI Fundamentals (AI-900) — cloud and AI services foundations for production deployments. Verify ↗

  • Azure cloud fundamentals
  • Azure AI services and responsible AI concepts

Google Cybersecurity Professional Certificate covering security operations, network security, and threat detection fundamentals. Verify ↗

  • Security operations and SIEM basics
  • Network security and threat detection

Contact

Let's work together

Open to internships, research collaborations, and full-stack AI roles. Email is the fastest way to reach me.

Send Email ↗ GitHub LinkedIn